Optics and Precision Engineering, Volume. 30, Issue 17, 2119(2022)

Neural architecture search algorithm based on voting scheme

Jun YANG1,2、* and Jingfa ZHANG1
Author Affiliations
  • 1School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou730070, China
  • 2Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou730070, China
  • show less
    Figures & Tables(15)
    Maurice Kendall coefficient in search and evaluation phases
    Overall network frame
    Comparison of sampling method
    Cell structure
    Optimal Cell structure obtained by proposed NAS-VS method
    Correlation of Maurice Kendall coefficient
    Classification accuracy of different sampling methods
    Weight change of each candidate operation
    [in Chinese]
    • Table 1. Comparison of recognition accuracy of different algorithms on ModelNet40

      View table
      View in Article

      Table 1. Comparison of recognition accuracy of different algorithms on ModelNet40

      网络架构分类准确率OA (%)参数(M)GPU 天数(GPU-days)
      KPConv2292.914.3人工设计
      A-CNN2392.62.05人工设计
      Point2Seq2492.62.67人工设计
      DGCNN2592.21.48人工设计
      Grid-GCN2693.18.2人工设计
      3DmFVNet2791.60.6人工设计
      Geo-CNN2893.412.07人工设计
      PAConv2993.69.07人工设计
      GS-Net3093.36.52人工设计
      SGAS1892.98.980.19
      DARTS691.812.360.36
      Noisy-DARTS2192.98.480.27
      NAS-VS(最高)93.99.020.24
      NAS-VS(平均)93.58.810.26
    • Table 2. Classification effect of NAS-VS on ModelNet40

      View table
      View in Article

      Table 2. Classification effect of NAS-VS on ModelNet40

      架构名称生成架构类型分类准确率OA(%)
      NAS-VS_1

      Genotype(normal=[('mr_conv',0), ('mr_conv',1), ('mr_conv',1), ('conv_1×1',2),

      ('skip_connect',0), ('gin',1)], normal_concat = range (1, 5))

      93.7
      NAS-VS_2

      Genotype(normal=[('edge_conv',0), ('skip_connect',1), ('conv_1×1', 0), ('conv_1×1',2),

      ('skip_connect',1), ('gat',2)], normal_concat=range (1, 5))

      93.0
      NAS-VS_3

      Genotype(normal=[('edge_conv',0), ('edge_conv',1), ('skip_connect',1), ('gin',2),

      ('mr_conv',1), ('edge_conv',2)], normal_concat=range (1, 5))

      93.7
      NAS-VS_4

      Genotype(normal=[('edge_conv',0), ('edge_conv',1), ('skip_connect',1), ('gat',2),

      ('conv_1×1',0), ('gat',3)], normal_concat=range (1, 5))

      93.4
      NAS-VS_5

      Genotype(normal=[('conv_1×1',0), ('mr_conv',1), ('edge_conv',0),('mr_conv',1),

      ('conv_1×1',0),('gat',1)], normal_concat=range (1, 5))

      93.9
    • Table 3. Weighted voting selection process for optimal Cell structure

      View table
      View in Article

      Table 3. Weighted voting selection process for optimal Cell structure

      节点对可微架构搜索策略噪声策略组稀疏正则化策略加权投票
      C_{k-1}-1mr_conv S=0.186mr_conv S=0.154mr_conv S=0.173mr_conv
      C_{k-1}-2

      mr_conv S=0.189

      semi_gcn S=0.190

      gat S=0.174mr_conv S=0.131mr_conv
      C_{k-1}-3skip_connect S=0.240gat S=0.153gat S=0.145gat
      C_{k-2}-1sage S=0.163conv_1×1 S=0.179semi_gcn S=0.143conv_1×1
      C_{k-2}-2skip_connect S=0.216

      edge_conv S=0.148

      mr_conv S=0.147

      edge_conv S=0.151edge_conv
      C_{k-2}-3semi_gcn S=0.128mr_conv S=0.157conv_1×1 S=0.164conv_1×1
    • Table 4. Influence of different sampling methods

      View table
      View in Article

      Table 4. Influence of different sampling methods

      采样方式分类准确率OA(%)
      均匀采样92.1
      性能估计器93.9
    • Table 5. Impact of skip connections on search space

      View table
      View in Article

      Table 5. Impact of skip connections on search space

      搜索空间分类准确率OA(%)
      Ω193.9
      Ω292.6
    • Table 6. Comparison of weighted voting experiments

      View table
      View in Article

      Table 6. Comparison of weighted voting experiments

      可微架构搜索策略噪声策略组稀疏正则化策略分类准确率OA(%)
      ×93.1
      ×92.5
      ×93.3
      93.9
    Tools

    Get Citation

    Copy Citation Text

    Jun YANG, Jingfa ZHANG. Neural architecture search algorithm based on voting scheme[J]. Optics and Precision Engineering, 2022, 30(17): 2119

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Information Sciences

    Received: Feb. 15, 2022

    Accepted: --

    Published Online: Oct. 20, 2022

    The Author Email: Jun YANG (yangj@mail.lzjtu.cn)

    DOI:10.37188/OPE.20223017.2119

    Topics